Crop Classification using Multi-spectral and Multitemporal Satellite Imagery with Machine Learning

Abstract

Satellite images are highly utilized for detecting land usage, while in recent years a finer-grade crop classification has become important in the context of precision agriculture. However, such classification brings new challenges, which aside from multi-spectral images require exploitation of their multi-temporal properties as well, with pixel-based analysis and larger number of classes. In this paper, we apply several machine learning algorithms on multi-spectral and multi-temporal satellite images and derive crop classification models. The models are applied only on agricultural fields, which can be singled out with the existing land usage classification models. Results show that the random forest outperforms other algorithms with accuracy score of 0.8420 and Kappa score of 0.8157. Detailed analysis of recall and precision scores is given for each crop separately, followed by a comprehensive discussion.